Background. Outline. Radiographic Tomosynthesis: Image Quality and Artifacts Reduction 1 / GE /

Transcription

1 Radiographic Tomosynthesis: Image Quality and Artifacts Reduction Baojun Li, Ph.D Department of Radiology Boston University Medical Center 2012 AAPM Annual Meeting Background Linear Trajectory Tomosynthesis Retrospective reconstruction of tomographic planes from multiple lowdose projections (21-61) acquired over a limited angular range (<60 ) along a linear trajectory. Outline Image quality metrics: methods and results Spatial resolution In-plane (x-y plane) Slice sensitivity profile (SSP) (z-dimension) Low contrast detectability Pulmonary nodule detection Artifacts and remedies Out of focus objects Ripples Edge fall off 1 /

2 Raw Image MTF 1,2,3,4 pixel Boxcar Measure MTF using a slanted wire (Flynn et. al, SPIE 2007) um tungsten wire suspended in air. Slant angle of 6 degree. Obtain point-spread-function in the scanning direction. MTF = FFT(PSF) normalized to 1. PSF Tungsten wire A tomo image PSF Mid-frequency hump shape of edge enhancing filters MTF(0) some vendors add a small DC component, others don t. MTF Tomo image of a spherical object Tomosynthesis vs. digital linear tomography MTF 0.1 ~= 1.3 lp/mm lp/mm MTF Tomosynthesis (150 um x 150 um) Linear Tomography (200 um x 200 um) 50% MTF ~1.8 lp/mm ~0.5 lp/mm 10% MTF ~3.5 lp/mm ~1.3 lp/mm The in-plane resolution of Tomosynthesis is 3x of digital linear tomography 2 /

3 Tomosynthesis vs. CT Azimuthally averaged MTF MTF50 = lp/cm, MTF10 = 18.15, MTF4 = VCT (standard kernel) lp/cm Focal spot wobble MTF Tomosynthesis MDCT (0.65 mm) MDCT (0.65 mm) (standard kernel) (150 um x 150 um) (FS wobble) 50% MTF ~1.8 lp/mm ~0.4 lp/mm ~1.2 lp/mm 10% MTF ~3.5 lp/mm ~0.7 lp/mm ~1.8 lp/mm The in-plane resolution of Tomosynthesis is 2~4x of CT Comparison of clinical images (Flynn et. al, SPIE 2007) CT (coronal reformat) Tomosynthesis (150 um x 150 um) Summary The in-plane resolution of tomosynthesis is ~3x of digital linear tomography The in-plane resolution of tomosynthesis is ~2-4x of CT 3 /

4 Slice Sensitivity Profile SSP is reconstruction algorithm dependent For Filtered backprojection, the relationship is derived (Li et. al, Med Phys 2008) θ / 2 D / 2 f ( x0, z0 ) = dϕ q( x, ϕ) h( x p ( ϕ, x0, z0 ) x ) dx θ / 2 D / 2 θ / 2 D / 2 SID tan( ϕ) x0 = d q x h x + z x dx ϕ (, ϕ) ( 0 0 ) SID z θ / 2 D / 2 0 f ( x0, z0) f ( x0, z0) f ( x0, z1 ) tan( ϑ / 2) ϑ System geometry for SSP derivation 1. For FBP, the SSP is inversely proportional to angular range θ. 2. SSP is spatially varying as a function of z 0. Slice Sensitivity Profile Simulation study LSF The reduction in SSP is less and less pronounced beyond 40 Slice Sensitivity Profile Anthropomorphic phantom (Li et. al, Med Phys 2008) (a) In-focus plane containing a 8.8 mm nodule (b) (d) Off-focus plane at z=110 mm with increasing angular range from 20 to 60 Confirming: A decrease in slice thickness as angular range increases Diminishing returns beyond 40 4 /

5 Slice Sensitivity Profile Measure MTF using a slanted wedge (Li et. al, Med Phys 2008) Tungsten ruler on a foam support. Slant angle of 30 degree Obtain a family of line-spread-functions in the scanning direction PSFs = Differentiate(LSFs) Obtain FWHMs from the family of PSFs as function of z PSF Tungsten wire PSF A tomo image Slice Sensitivity Profile Tomosynthesis vs. CT SSP Tomosynthesis MDCT MDCT (zffs) (200 um x 200 um) FWHM ~ 3 mm ~ 0.6 mm ~ 0.3 mm The slice thickness of Tomosynthesis is ~ 5-10x of CT Slice Sensitivity Profile Summary For FBP, the SSP is inversely proportional to angular range θ; diminishing returns beyond 40 The slice thickness of Tomosynthesis is ~5-10x of CT i.e., highly anisotropic voxel SSP of tomosynthesis is spatially varying 5 /

6 Low Contrast Detectability Simulated pulmonary nodule DES vs DES-Tomosynthesis (Gomi et. al, Acad. Radiol. 2011) The area under ROC curve was ~90% and ~60% for DES-T and DES, respectively (P <.003). Low Contrast Detectability Nodule detection clinical trial (Dobbins et. al, Med Phys 2008) Low Contrast Detectability Summary Low contrast detectability of Tomosynthesis is ~3x (overall) of PA radiography Low contrast detectability of Tomosynthesis is sizedependent: 0.5x of CT for <5mm nodules, 0.7x of CT for 5-10mm nodules, and comparable for > 10mm nodules 6 /

7 Out of focus objects Highly anisotropic voxel size due to limited angular range (<60 ) Objects appear in adjacent slices where they do not belong Out of focus objects lead to blurry image and anatomical noise Z Y X Highly anisotropic voxel size Out of focus objects Out of focus objects Artifacts can be suppressed by software Order statistics based approach (Claus et. al, SPIE 2002) Before After Out of focus objects Hybrid approach combining order statistics based method with advanced image segmentation (Wu et. al, Med Phys 2006) Before After 7 /

8 Ripples Limited # of projection results in under-sampling of the angular range Discrete impulse response is a result of angular aliasing Strong edges (e.g., ribs, implants, etc.) manifests as ripples Impulse response from N = 11 Ripples observed in phantom Images (thorax) Ripples Artifacts can be suppressed by optimization of acquisition parameters or reconstruction algorithm (Dobbins et. al, Med Phys 2008) N = 11 N = 11, MITS N = 61 N = 61, MITS Ripples Artifacts can be suppressed by optimization of acquisition (Deller et. al, SPIE 2007) N = 21, FBP N = 61, FBP 8 /

9 Edge fall off Uneven number of rays passing thru the planes due to truncation Truncated projections cause intensity drop off Before For rad tomo, source-to-image distance (SID) around 100 cm (table) or 180 cm (wall-stand) Edge fall off Artifacts can be corrected by software 3-D non-uniform view-weighting technique (Li et. al, SPIE 2007) Iterative local intensity equalization method (Zhang et. al, JCAR 2009) 64% - 77% 80% - 89% Before After Conclusion Image quality metrics of linear-trajectory Tomo In-plane resolution: 2-4x of CT or linear tomography SSP: inversely proportional to θ; slice thickness: 5-10x thicker than CT Low contrast detectability: >3x better than radiography, but somewhat inferior to CT (nodule size-dependent) Image artifacts and reduction strategies Out of focus objects post the biggest image quality challenge; Order statistics based method helps, but is not perfect yet Ripple artifacts can be suppressed sufficiently by optimization of acquisition parameters and/or reconstruction algorithm Edge fall off can be corrected by 3-D non-uniform view-weighting or iterative intensity equalization 9 /